Merge branch 'main' into dr-est_snr

CMakeLists.txt

@@ -251,7 +251,7 @@ add_test(NAME radae_tx_basic
                        ./inference.sh model19_check3/checkpoints/checkpoint_epoch_100.pth wav/brian_g8sez.wav /dev/null \
                        --rate_Fs --pilots --pilot_eq --eq_ls --cp 0.004 --bottleneck 3 --auxdata --write_rx rx.f32 --correct_freq_offset; \
                        cat features_in.f32 | python3 radae_txe.py --model model19_check3/checkpoints/checkpoint_epoch_100.pth --txbpf > rx.f32
-                       cat rx.f32 | python3 radae_rxe.py --model model19_check3/checkpoints/checkpoint_epoch_100.pth -v 1 > features_txs_out.f32; \
+                       cat rx.f32 | python3 radae_rxe.py --model model19_check3/checkpoints/checkpoint_epoch_100.pth -v 2 > features_txs_out.f32; \
                        python3 loss.py features_in.f32 features_txs_out.f32 --loss_test 0.15 --acq_time_test 0.5 --clip_start 5")
                        set_tests_properties(radae_tx_basic PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
 
@@ -291,10 +291,10 @@ add_test(NAME rx_streaming
 # basic test of streaming rx, run rx in vanilla and streaming, compare
 add_test(NAME radae_rx_basic
         COMMAND sh -c "cd ${CMAKE_SOURCE_DIR}; \
-                       ./inference.sh model17/checkpoints/checkpoint_epoch_100.pth wav/brian_g8sez.wav /dev/null \
-                       --EbNodB 10 --freq_offset 11 \
+                       ./inference.sh model19_check3/checkpoints/checkpoint_epoch_100.pth wav/brian_g8sez.wav /dev/null \
+                       --EbNodB 10 --freq_offset 11 --prepend_noise 1 --append_noise 1 --end_of_over --auxdata \
                        --rate_Fs --pilots --pilot_eq --eq_ls --cp 0.004 --bottleneck 3 --write_rx rx.f32 --correct_freq_offset; \
-                       cat rx.f32 | PYTHONPATH='../' python3 radae_rxe.py --model model17/checkpoints/checkpoint_epoch_100.pth -v 1 --noauxdata > features_rxs_out.f32; \
+                       cat rx.f32 | PYTHONPATH='../' python3 radae_rxe.py -v 2 > features_rxs_out.f32; \
                        python3 loss.py features_in.f32 features_rxs_out.f32 --loss_test 0.15 --acq_time_test 0.5")
                        set_tests_properties(radae_rx_basic PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
 
@@ -309,7 +309,7 @@ add_test(NAME radae_rx_awgn
                        --rate_Fs --pilots --pilot_eq --eq_ls --cp 0.004 --bottleneck 3 --time_offset -16 --write_rx rx.f32  \
                        --prepend_noise 1 --append_noise 3 --end_of_over --auxdata --correct_freq_offset; \
                        cat rx.f32 | python3 radae_rxe.py --model model19_check3/checkpoints/checkpoint_epoch_100.pth -v 2 > features_rx_out.f32; \
-                       python3 loss.py features_in.f32 features_rx_out.f32 --loss 0.3 --acq_time_test 1.0 --clip_end 100")
+                       python3 loss.py features_in.f32 features_rx_out.f32 --loss 0.3 --acq_time_test 1.0 --clip_end 300")
                        set_tests_properties(radae_rx_awgn PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
 
 # SNR=0dB MPP
@@ -488,7 +488,8 @@ add_test(NAME radae_tx_embed_c
 add_test(NAME radae_rx_embed_c
         COMMAND sh -c "cd ${CMAKE_SOURCE_DIR}; \
                        ./inference.sh model19_check3/checkpoints/checkpoint_epoch_100.pth wav/brian_g8sez.wav /dev/null \
-                       --rate_Fs --pilots --pilot_eq --eq_ls --cp 0.004 --bottleneck 3 --auxdata --write_rx rx.f32 --correct_freq_offset; \
+                       --EbNodB 100 --freq_offset 0 --append_noise 1 --end_of_over --auxdata \
+                       --rate_Fs --pilots --pilot_eq --eq_ls --cp 0.004 --bottleneck 3 --write_rx rx.f32 --correct_freq_offset; \
                        cat rx.f32 | PYTHONPATH='.' ${CMAKE_CURRENT_BINARY_DIR}/src/radae_rx > features_out.f32;
                        python3 loss.py features_in.f32 features_out.f32 --loss_test 0.15 --acq_time_test 0.5")
                        set_tests_properties(radae_rx_embed_c PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
@@ -551,6 +552,41 @@ add_test(NAME c_decoder_aux_mpp
                        python3 loss.py features_in.f32 features_c.f32 --loss 0.3 --clip_start 300")
                        set_tests_properties(c_decoder_aux_mpp PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
 
+# EOO data -------------------------------------------------------------------------------------------
+
+# Pythion Tx & Rx, no noise
+add_test(NAME radae_eoo_data_py
+        COMMAND sh -c "cd ${CMAKE_SOURCE_DIR}; \
+                       ${CMAKE_CURRENT_BINARY_DIR}/src/lpcnet_demo -features wav/brian_g8sez.wav features_in.f32; \
+                       cat features_in.f32 | python3 radae_txe.py --eoo_data_test > rx.f32; \
+                       cat rx.f32 | python3 radae_rxe.py -v 2 --eoo_data_test > /dev/null")
+                       set_tests_properties(radae_eoo_data_py PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
+
+# C tx & rx, no noise. Note Python radae_txe.py just generates eoo_tx.f32 for C radae_tx
+add_test(NAME radae_eoo_data_c
+        COMMAND sh -c "cd ${CMAKE_SOURCE_DIR}; \
+                       ${CMAKE_CURRENT_BINARY_DIR}/src/lpcnet_demo -features wav/brian_g8sez.wav features_in.f32; \
+                       cat features_in.f32 | python3 radae_txe.py --eoo_data_test > /dev/null; \
+                       cat features_in.f32 | PYTHONPATH='.' ${CMAKE_CURRENT_BINARY_DIR}/src/radae_tx > rx.f32; \
+                       cat rx.f32 | PYTHONPATH='.' ${CMAKE_CURRENT_BINARY_DIR}/src/radae_rx > /dev/null; \
+                       python3 eoo_ber.py eoo_tx.f32 eoo_rx.f32")
+                       set_tests_properties(radae_eoo_data_c PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
+
+# C tx & rx, over a multipath channel, we set up 5 "overs", each with an EOO chunk of data.  Just one of them needs
+# to have a BER < 5% for a PASS.  About 6dB SNR
+add_test(NAME radae_eoo_data_mpp
+        COMMAND sh -c "cd ${CMAKE_SOURCE_DIR}; test/make_g.sh; \
+                       ${CMAKE_CURRENT_BINARY_DIR}/src/lpcnet_demo -features wav/brian_g8sez.wav features_in.f32; \
+                       cat features_in.f32 | python3 radae_txe.py --eoo_data_test > /dev/null; \
+                       cat features_in.f32 | PYTHONPATH='.' ${CMAKE_CURRENT_BINARY_DIR}/src/radae_tx > tx.f32; \
+                       cat tx.f32 tx.f32 tx.f32 tx.f32 tx.f32 > tx_2.f32;
+                       cat tx_2.f32 | python3 f32toint16.py --real --scale 8192 > tx.raw; \
+                       ${CODEC2_DEV_BUILD_DIR}/src/ch tx.raw rx.raw --No -24 --mpp --fading_dir .; \
+                       cat rx.raw | python3 int16tof32.py --zeropad > rx.f32; \
+                       cat rx.f32 | PYTHONPATH='.' ${CMAKE_CURRENT_BINARY_DIR}/src/radae_rx > /dev/null; \
+                       python3 eoo_ber.py eoo_tx.f32 eoo_rx.f32")
+                       set_tests_properties(radae_eoo_data_mpp PROPERTIES PASS_REGULAR_EXPRESSION "PASS")
+
 # BBFM -----------------------------------------------------------------------------------------------
 
 # single carrier modem internal (inside single_carrier class) tests

README.md

@@ -51,7 +51,6 @@ The RDOVAE derived Python source code is released under the two-clause BSD licen
 | stateful_encoder.[py,sh] | Inference test that compares stateful to vanilla encoder |
 | radae_tx.[py,sh] | streaming RADAE encoder and helper script |
 | radae_rx.[py,sh] | streaming RADAE decoder and helper script |
-| resource_est.py | WIP estimate CPU/memory resources |
 | radae_base.py | Shared ML code between models |
 | radae/bbfm.py | Baseband FM PyTorch model |
 | train_bbfm.py | Training for BBFM model |

eoo_ber.py

@@ -0,0 +1,18 @@
+import sys
+import numpy as np
+
+# bits are in float form, e.g. +/-1 or +/-1000
+tx_bits = np.fromfile(sys.argv[1], dtype=np.float32)
+rx_bits = np.fromfile(sys.argv[2], dtype=np.float32)
+bits_per_frame = len(tx_bits)
+n_frames = len(rx_bits)//bits_per_frame
+n_ok_frames = 0
+for f in range(n_frames):
+    n_errors = sum(rx_bits[f*bits_per_frame:(f+1)*bits_per_frame]*tx_bits < 0)
+    ber = n_errors/bits_per_frame
+    print(f"frame received! BER: {ber:5.2f}")
+    if ber < 0.05:
+        n_ok_frames += 1
+print(f"EOO frames  received: {n_frames} n_ok_frames: {n_ok_frames}", file=sys.stderr)
+if n_ok_frames:
+    print("PASS", file=sys.stderr)

inference.py

@@ -57,7 +57,7 @@
 parser.add_argument('--rx_gain', type=float, default=1.0, help='gain to apply to --write_rx samples (default 1.0)')
 parser.add_argument('--write_tx', type=str, default="", help='path to output file of rate Fs tx samples in ..IQIQ...f32 format')
 parser.add_argument('--phase_offset', type=float, default=0, help='phase offset in rads')
-parser.add_argument('--freq_offset', type=float, help='freq offset in Hz')
+parser.add_argument('--freq_offset', type=float, default=0, help='freq offset in Hz')
 parser.add_argument('--time_offset', type=int, default=0, help='sampling time offset in samples')
 parser.add_argument('--df_dt', type=float, default=0, help='rate of change of freq offset in Hz/s')
 parser.add_argument('--gain', type=float, default=1.0, help='rx gain (defaul 1.0)')
@@ -240,7 +240,7 @@
       n_errors = int(torch.sum(x < 0))
       n_bits = int(torch.numel(x))
       BER = n_errors/n_bits
-      print(f"loss: {loss:5.3f} BER: {BER:5.3f}")
+      print(f"loss: {loss:5.3f} Auxdata BER: {BER:5.3f}")
    else:
       print(f"loss: {loss:5.3f}")
    if args.loss_test > 0.0:
@@ -264,6 +264,14 @@
             # appends a frame containing a final pilot so the last RADAE frame
             # has a good phase reference, and two "end of over" symbols
             eoo = model.eoo
+
+            # this is messy! - continue phase, freq and dF/dt track from inside forward()
+            freq = torch.zeros_like(eoo)
+            freq[:,] = model.freq_offset*torch.ones_like(eoo) + model.df_dt*torch.arange(eoo.shape[1])/model.Fs
+            omega = freq*2*torch.pi/model.Fs
+            lin_phase = torch.cumsum(omega,dim=1)
+            lin_phase = torch.exp(1j*lin_phase)
+            eoo = eoo*lin_phase*model.final_phase
             eoo = eoo + sigma*torch.randn_like(eoo)
             rx = torch.concatenate([rx,eoo],dim=1)
          if args.prepend_noise > 0.0:

radae/dsp.py

@@ -344,7 +344,7 @@ def transmitter_one(self, z, num_timesteps_at_rate_Rs):
 
 # Single modem frame streaming receiver. TODO: is there a better way to pass a bunch of constants around?
 class receiver_one():
-   def __init__(self,latent_dim,Fs,M,Ncp,Wfwd,Nc,Ns,w,P,bottleneck,pilot_gain,time_offset,coarse_mag):
+   def __init__(self,latent_dim,Fs,M,Ncp,Wfwd,Nc,Ns,w,P,Pend,bottleneck,pilot_gain,time_offset,coarse_mag):
       self.latent_dim = latent_dim
       self.Fs = Fs
       self.M = M
@@ -354,6 +354,7 @@ def __init__(self,latent_dim,Fs,M,Ncp,Wfwd,Nc,Ns,w,P,bottleneck,pilot_gain,time_
       self.Ns = Ns
       self.w = w
       self.P = P
+      self.Pend = Pend
       self.bottleneck = bottleneck
       self.pilot_gain = pilot_gain
       self.time_offset = time_offset
@@ -446,7 +447,7 @@ def do_pilot_eq_one(self, num_modem_frames, rx_sym_pilots):
       return rx_sym_pilots
 
    #  One frame version of rate Fs receiver for streaming implementation
-   def receiver_one(self, rx):
+   def receiver_one(self, rx, endofover):
       Ns = self.Ns + 1
 
       # we expect: Pilots - data symbols - Pilots
@@ -462,21 +463,32 @@ def receiver_one(self, rx):
       # DFT to transform M time domain samples to Nc carriers
       rx_sym = torch.matmul(rx_dash, self.Wfwd)
 
-      # Pilot based EQ
       rx_sym_pilots = torch.reshape(rx_sym,(1, num_modem_frames, num_timesteps_at_rate_Rs, self.Nc))
-      rx_sym_pilots = self.do_pilot_eq_one(num_modem_frames,rx_sym_pilots)
-      rx_sym = torch.ones(1, num_modem_frames, self.Ns, self.Nc, dtype=torch.complex64)
-      rx_sym = rx_sym_pilots[:,:,1:self.Ns+1,:]
-
-      # demap QPSK symbols
-      rx_sym = torch.reshape(rx_sym, (1, -1, self.latent_dim//2))
-      z_hat = torch.zeros(1,rx_sym.shape[1], self.latent_dim)
-
-      z_hat[:,:,::2] = rx_sym.real
-      z_hat[:,:,1::2] = rx_sym.imag
-
+      if not endofover:
+         # Pilot based least squares EQ
+         rx_sym_pilots = self.do_pilot_eq_one(num_modem_frames,rx_sym_pilots)
+         rx_sym = rx_sym_pilots[:,:,1:self.Ns+1,:]
+         rx_sym = torch.reshape(rx_sym, (1, -1, self.latent_dim//2))
+         z_hat = torch.zeros(1,rx_sym.shape[1], self.latent_dim)
+
+         z_hat[:,:,::2] = rx_sym.real
+         z_hat[:,:,1::2] = rx_sym.imag
+      else:
+         # Simpler (but lower performance) EQ as average of pilots, as LS set up for PDDDDP, rather than out PEDDDE
+         for c in range(self.Nc):
+            phase_offset = torch.angle(rx_sym_pilots[0,0,0,c]/self.P[c] +
+                           rx_sym_pilots[0,0,1,c]/self.Pend[c] +
+                           rx_sym_pilots[0,0,Ns,c]/self.Pend[c])
+            rx_sym_pilots[:,:,:Ns+1,c] *= torch.exp(-1j*phase_offset)
+         rx_sym = torch.reshape(rx_sym_pilots[:,:,2:Ns,:],(1,(Ns-2)*self.Nc))
+         z_hat = torch.zeros(1,(Ns-2)*self.Nc*2)
+
+         z_hat[:,::2] = rx_sym.real
+         z_hat[:,1::2] = rx_sym.imag
+
       return z_hat
 
+
 # Generate root raised cosine (Root Nyquist) filter coefficients
 # thanks http://www.dsplog.com/db-install/wp-content/uploads/2008/05/raised_cosine_filter.m
 

radae/radae.py

@@ -199,12 +199,12 @@ def __init__(self,
             self.pilot_gain = pilot_backoff*self.M/(Nc**0.5)
 
         self.d_samples = int(self.multipath_delay * self.Fs)         # multipath delay in samples
-        self.Ncp = int(cyclic_prefix*self.Fs)
 
         # set up End Of Over sequence
         # Normal frame ...PDDDDP... 
         # EOO frame    ...PE000E... 
         # Key: P = self.p_cp, D = data symbols, E = self.pend_cp, 0 = zeros
+
         if self.Ncp:
             M = self.M
             Ncp = self.Ncp
@@ -217,7 +217,10 @@ def __init__(self,
             if self.bottleneck == 3:
                 eoo = torch.tanh(torch.abs(eoo)) * torch.exp(1j*torch.angle(eoo))
             self.eoo = eoo
-
+
+        # experimental EOO data symbols (quick and dirty supplimentary txt channel)
+        self.Nseoo = (Ns-1)*Nc  # number of EOO data symbols
+
         print(f"Rs: {Rs:5.2f} Rs': {Rs_dash:5.2f} Ts': {Ts_dash:5.3f} Nsmf: {Nsmf:3d} Ns: {Ns:3d} Nc: {Nc:3d} M: {self.M:d} Ncp: {self.Ncp:d}", file=sys.stderr)
 
         self.Tmf = Tmf
@@ -435,6 +438,22 @@ def est_snr(self, r, time_offset=0):
         SNR_est = Ct/(torch.dot(torch.conj(p),p) - Ct)
         return SNR_est.real
 
+    def set_eoo_bits(self, eoo_bits):
+        Ns = self.Ns; Ncp = self.Ncp; M = self.M; Nc = self.Nc; Nmf = int((Ns+1)*(M+Ncp))
+
+        eoo_syms = eoo_bits[::2] + 1j*eoo_bits[1::2]
+        eoo_syms = torch.reshape(eoo_syms,(1,Ns-1,Nc))
+
+        eoo_tx = torch.matmul(eoo_syms,self.Winv)
+        if self.Ncp:
+            eoo_tx_cp = torch.zeros((1,Ns-1,self.M+Ncp),dtype=torch.complex64)
+            eoo_tx_cp[:,:,Ncp:] = eoo_tx
+            eoo_tx_cp[:,:,:Ncp] = eoo_tx_cp[:,:,-Ncp:]
+            eoo_tx = torch.reshape(eoo_tx_cp,(1,(Ns-1)*(self.M+Ncp)))*self.pilot_gain
+            if self.bottleneck == 3:
+                eoo_tx = torch.tanh(torch.abs(eoo_tx)) * torch.exp(1j*torch.angle(eoo_tx))
+            self.eoo[0,2*(M+Ncp):Nmf] = eoo_tx
+
     def forward(self, features, H, G=None):
 
         (num_batches, num_ten_ms_timesteps, num_features) = features.shape
@@ -482,6 +501,7 @@ def forward(self, features, H, G=None):
 
         tx_before_channel = None
         rx = None
+        self.final_phase = torch.tensor(1,dtype=torch.complex64)
         if self.rate_Fs:
             num_timesteps_at_rate_Fs = num_timesteps_at_rate_Rs*self.M
 
@@ -530,6 +550,7 @@ def forward(self, features, H, G=None):
                 lin_phase = torch.cumsum(omega,dim=1)
                 lin_phase = torch.exp(1j*lin_phase)
                 tx = tx*lin_phase
+                self.final_phase = lin_phase[:,-1]
 
             # insert per sequence random phase and freq offset (training time)
             if self.freq_rand:

radae/radae_base.py

@@ -185,7 +185,7 @@ def __init__(self, feature_dim, output_dim, bottleneck = 1):
         self.z_dense = nn.Linear(864, self.output_dim)
 
         nb_params = sum(p.numel() for p in self.parameters())
-        print(f"encoder: {nb_params} weights", file=sys.stderr)
+        #print(f"encoder: {nb_params} weights", file=sys.stderr)
 
         # initialize weights
         self.apply(init_weights)
@@ -251,7 +251,7 @@ def __init__(self, feature_dim, output_dim, bottleneck = 1):
         self.z_dense = nn.Linear(864, self.output_dim)
 
         nb_params = sum(p.numel() for p in self.parameters())
-        print(f"encoder: {nb_params} weights", file=sys.stderr)
+        #print(f"encoder: {nb_params} weights", file=sys.stderr)
 
         # initialize weights
         self.apply(init_weights)
@@ -326,7 +326,7 @@ def __init__(self, input_dim, output_dim):
         self.glu5 = GLU(96)
 
         nb_params = sum(p.numel() for p in self.parameters())
-        print(f"decoder: {nb_params} weights", file=sys.stderr)
+        #print(f"decoder: {nb_params} weights", file=sys.stderr)
         # initialize weights
         self.apply(init_weights)
 
@@ -393,7 +393,7 @@ def __init__(self, input_dim, output_dim):
         self.glu5 = GLU(96)
 
         nb_params = sum(p.numel() for p in self.parameters())
-        print(f"decoder: {nb_params} weights", file=sys.stderr)
+        #print(f"decoder: {nb_params} weights", file=sys.stderr)
         # initialize weights
         self.apply(init_weights)