def collate_fn_checker():
dataBatch = list()
inpLens = [10, 8, 7, 10]
trgtLens = [4, 6, 7, 10]
for i in range(len(inpLens)):
inp = torch.from_numpy(np.random.rand(inpLens[i], 1, args["ROI_SIZE"], args["ROI_SIZE"]))
trgt = torch.from_numpy(np.random.randint(0, args["NUM_CLASSES"], trgtLens[i]))
inpLen = torch.tensor(inpLens[i])
trgtLen = torch.tensor(trgtLens[i])
data = (inp, trgt, inpLen, trgtLen)
dataBatch.append(data)
inputBatch, targetBatch, inputLenBatch, targetLenBatch = collate_fn(dataBatch)
print(inputBatch.shape, targetBatch.shape, inputLenBatch.shape, targetLenBatch.shape)
return
def collate_fn_checker():
dataBatch = list()
inpLens = [10, 8, 7, 10]
trgtLens = [4, 6, 7, 10]
for i in range(len(inpLens)):
audInp = torch.from_numpy(
np.random.rand(4 * inpLens[i], args["AUDIO_FEATURE_SIZE"]))
vidInp = torch.from_numpy(
np.random.rand(inpLens[i], args["TX_NUM_FEATURES"]))
inp = (audInp, vidInp)
trgt = torch.from_numpy(
np.random.randint(0, args["NUM_CLASSES"], trgtLens[i]))
inpLen = torch.tensor(inpLens[i])
trgtLen = torch.tensor(trgtLens[i])
data = (inp, trgt, inpLen, trgtLen)
dataBatch.append(data)
inputBatch, targetBatch, inputLenBatch, targetLenBatch = collate_fn(
dataBatch)
print((inputBatch[0].shape, inputBatch[1].shape), targetBatch.shape,
inputLenBatch.shape, targetLenBatch.shape)
return
#walking through the demo directory and running the model on all video files in it
for root, dirs, files in os.walk(args["DEMO_DIRECTORY"]):
for file in files:
if file.endswith(".mp4"):
sampleFile = os.path.join(root, file[:-4])
targetFile = os.path.join(root, file[:-4]) + ".txt"
#preprocessing the sample
preprocess_sample(sampleFile)
#converting the data sample into appropriate tensors for input to the model
audioFile = os.path.join(root, file[:-4]) + ".wav"
audioParams = {"stftWindow":args["STFT_WINDOW"], "stftWinLen":args["STFT_WIN_LENGTH"], "stftOverlap":args["STFT_OVERLAP"]}
inp, trgt, inpLen, trgtLen = prepare_main_input(audioFile, targetFile, noise, args["MAIN_REQ_INPUT_LENGTH"],
args["CHAR_TO_INDEX"], args["NOISE_SNR_DB"], audioParams)
inputBatch, targetBatch, inputLenBatch, targetLenBatch = collate_fn([(inp, trgt, inpLen, trgtLen)])
#running the model
inputBatch, targetBatch = (inputBatch.float()).to(device), (targetBatch.int()).to(device)
inputLenBatch, targetLenBatch = (inputLenBatch.int()).to(device), (targetLenBatch.int()).to(device)
model.eval()
with torch.no_grad():
outputBatch = model(inputBatch)
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(outputBatch, inputLenBatch, args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
beamSearchParams = {"beamWidth":args["BEAM_WIDTH"], "alpha":args["LM_WEIGHT_ALPHA"], "beta":args["LENGTH_PENALTY_BETA"],
"threshProb":args["THRESH_PROBABILITY"]}
#preprocessing the sample
params = {
"roiSize": args["ROI_SIZE"],
"normMean": args["NORMALIZATION_MEAN"],
"normStd": args["NORMALIZATION_STD"],
"vf": vf
}
preprocess_sample(sampleFile, params)
#converting the data sample into appropriate tensors for input to the model
visualFeaturesFile = os.path.join(root, file[:-4]) + ".npy"
videoParams = {"videoFPS": args["VIDEO_FPS"]}
inp, _, inpLen, _ = prepare_main_input(
visualFeaturesFile, None, args["MAIN_REQ_INPUT_LENGTH"],
args["CHAR_TO_INDEX"], videoParams)
inputBatch, _, inputLenBatch, _ = collate_fn([(inp, None,
inpLen, None)])
#running the model
inputBatch = (inputBatch.float()).to(device)
inputLenBatch = (inputLenBatch.int()).to(device)
model.eval()
with torch.no_grad():
outputBatch = model(inputBatch)
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(
outputBatch, inputLenBatch,
args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
def main():
np.random.seed(args["SEED"])
torch.manual_seed(args["SEED"])
gpuAvailable = torch.cuda.is_available()
device = torch.device("cuda" if gpuAvailable else "cpu")
if args["TRAINED_MODEL_FILE"] is not None:
print("\nTrained Model File: %s" % (args["TRAINED_MODEL_FILE"]))
print("\nDemo Directory: %s" % (args["DEMO_DIRECTORY"]))
#declaring the model and loading the trained weights
model = VideoNet(args["TX_NUM_FEATURES"], args["TX_ATTENTION_HEADS"],
args["TX_NUM_LAYERS"], args["PE_MAX_LENGTH"],
args["TX_FEEDFORWARD_DIM"], args["TX_DROPOUT"],
args["NUM_CLASSES"])
model.load_state_dict(
torch.load(args["CODE_DIRECTORY"] + args["TRAINED_MODEL_FILE"],
map_location=device))
model.to(device)
#declaring the visual frontend module
vf = VisualFrontend()
vf.load_state_dict(
torch.load(args["TRAINED_FRONTEND_FILE"], map_location=device))
vf.to(device)
#declaring the language model
lm = LRS2CharLM()
lm.load_state_dict(
torch.load(args["TRAINED_LM_FILE"], map_location=device))
lm.to(device)
if not args["USE_LM"]:
lm = None
print("\n\nRunning Demo .... \n")
#walking through the demo directory and running the model on all video files in it
for root, dirs, files in os.walk(args["DEMO_DIRECTORY"]):
for file in files:
if file.endswith(".mp4"):
sampleFile = os.path.join(root, file[:-4])
#preprocessing the sample
params = {
"roiSize": args["ROI_SIZE"],
"normMean": args["NORMALIZATION_MEAN"],
"normStd": args["NORMALIZATION_STD"],
"vf": vf
}
preprocess_sample(sampleFile, params)
#converting the data sample into appropriate tensors for input to the model
visualFeaturesFile = os.path.join(root, file[:-4]) + ".npy"
videoParams = {"videoFPS": args["VIDEO_FPS"]}
inp, _, inpLen, _ = prepare_main_input(
visualFeaturesFile, None,
args["MAIN_REQ_INPUT_LENGTH"], args["CHAR_TO_INDEX"],
videoParams)
inputBatch, _, inputLenBatch, _ = collate_fn([
(inp, None, inpLen, None)
])
#running the model
inputBatch = (inputBatch.float()).to(device)
inputLenBatch = (inputLenBatch.int()).to(device)
model.eval()
with torch.no_grad():
outputBatch = model(inputBatch)
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(
outputBatch, inputLenBatch,
args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
beamSearchParams = {
"beamWidth": args["BEAM_WIDTH"],
"alpha": args["LM_WEIGHT_ALPHA"],
"beta": args["LENGTH_PENALTY_BETA"],
"threshProb": args["THRESH_PROBABILITY"]
}
predictionBatch, predictionLenBatch = ctc_search_decode(
outputBatch, inputLenBatch, beamSearchParams,
args["CHAR_TO_INDEX"][" "],
args["CHAR_TO_INDEX"]["<EOS>"], lm)
else:
print("Invalid Decode Scheme")
exit()
#converting character indices back to characters
pred = predictionBatch[:][:-1]
pred = "".join(
[args["INDEX_TO_CHAR"][ix] for ix in pred.tolist()])
#printing the predictions
print("File: %s" % (file))
print("Prediction: %s" % (pred))
print("\n")
print("Demo Completed.\n")
else:
print("\nPath to trained model file not specified.\n")
return
def main():
np.random.seed(args["SEED"])
torch.manual_seed(args["SEED"])
gpuAvailable = torch.cuda.is_available()
device = torch.device("cuda" if gpuAvailable else "cpu")
if args["TRAINED_MODEL_FILE"] is not None:
print("\nTrained Model File: %s" % (args["TRAINED_MODEL_FILE"]))
print("\nDemo Directory: %s" % (args["DEMO_DIRECTORY"]))
#declaring the model and loading the trained weights
model = AVNet(args["TX_NUM_FEATURES"], args["TX_ATTENTION_HEADS"],
args["TX_NUM_LAYERS"], args["PE_MAX_LENGTH"],
args["AUDIO_FEATURE_SIZE"], args["TX_FEEDFORWARD_DIM"],
args["TX_DROPOUT"], args["NUM_CLASSES"])
model.load_state_dict(
torch.load(args["CODE_DIRECTORY"] + args["TRAINED_MODEL_FILE"],
map_location=device))
model.to(device)
#declaring the visual frontend module
vf = VisualFrontend()
vf.load_state_dict(
torch.load(args["TRAINED_FRONTEND_FILE"], map_location=device))
vf.to(device)
#declaring the language model
lm = LRS2CharLM()
lm.load_state_dict(
torch.load(args["TRAINED_LM_FILE"], map_location=device))
lm.to(device)
if not args["USE_LM"]:
lm = None
#reading the noise file
if args["TEST_DEMO_NOISY"]:
_, noise = wavfile.read(args["DATA_DIRECTORY"] + "/noise.wav")
else:
noise = None
print("\n\nRunning Demo .... \n")
rows = []
print(args['TEST_DEMO_MODE'])
#walking through the demo directory and running the model on all video files in it
for root, dirs, files in os.walk(args["DEMO_DIRECTORY"]):
for file in files:
if file.endswith(".mp4"):
sNum = file[7]
if file[13] <= "9" and file[13] >= "0":
cNum = file[12:14]
else:
cNum = file[12]
if file[-6] == "l":
cType = "jumble"
elif file[-6] == "s":
cType = "base"
else:
cType = file[-8:-4]
sampleFile = os.path.join(root, file[:-4])
#preprocessing the sample
params = {
"roiSize": args["ROI_SIZE"],
"normMean": args["NORMALIZATION_MEAN"],
"normStd": args["NORMALIZATION_STD"],
"vf": vf
}
preprocess_sample(sampleFile, params)
#converting the data sample into appropriate tensors for input to the model
audioFile = os.path.join(root, file[:-4]) + ".wav"
visualFeaturesFile = os.path.join(root, file[:-4]) + ".npy"
audioParams = {
"stftWindow": args["STFT_WINDOW"],
"stftWinLen": args["STFT_WIN_LENGTH"],
"stftOverlap": args["STFT_OVERLAP"]
}
videoParams = {"videoFPS": args["VIDEO_FPS"]}
inp, _, inpLen, _ = prepare_main_input(
audioFile, visualFeaturesFile, None, noise,
args["MAIN_REQ_INPUT_LENGTH"], args["CHAR_TO_INDEX"],
args["NOISE_SNR_DB"], audioParams, videoParams)
inputBatch, _, inputLenBatch, _ = collate_fn([
(inp, None, inpLen, None)
])
#running the model
inputBatch = ((inputBatch[0].float()).to(device),
(inputBatch[1].float()).to(device))
inputLenBatch = (inputLenBatch.int()).to(device)
if args["TEST_DEMO_MODE"] == "AO":
inputBatch = (inputBatch[0], None)
elif args["TEST_DEMO_MODE"] == "VO":
inputBatch = (None, inputBatch[1])
elif args["TEST_DEMO_MODE"] == "AV":
pass
else:
print("Invalid Operation Mode.")
exit()
model.eval()
with torch.no_grad():
outputBatch = model(inputBatch)
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(
outputBatch, inputLenBatch,
args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
beamSearchParams = {
"beamWidth": args["BEAM_WIDTH"],
"alpha": args["LM_WEIGHT_ALPHA"],
"beta": args["LENGTH_PENALTY_BETA"],
"threshProb": args["THRESH_PROBABILITY"]
}
predictionBatch, predictionLenBatch = ctc_search_decode(
outputBatch, inputLenBatch, beamSearchParams,
args["CHAR_TO_INDEX"][" "],
args["CHAR_TO_INDEX"]["<EOS>"], lm)
else:
print("Invalid Decode Scheme")
exit()
#converting character indices back to characters
pred = predictionBatch[:][:-1]
pred = "".join(
[args["INDEX_TO_CHAR"][ix] for ix in pred.tolist()])
#printing the predictions
print("File: %s" % (file))
print("Speaker: " + sNum + " Clip: " + cNum +
" Clip Type: " + cType)
print("Prediction: %s" % (pred))
print("\n")
row = [sNum, cNum, cType, pred]
rows.append(row)
print("Demo Completed.\n")
with open("predictions.csv", "w", newline="") as file:
writer = csv.writer(file)
for x in range(len(rows)):
writer.writerow(rows[x])
else:
print("\nPath to trained model file not specified.\n")
return
def inference(sampleFile, targetFile=None):
print('In Inference')
#preprocessing the sample
params = {
"roiSize": args["ROI_SIZE"],
"normMean": args["NORMALIZATION_MEAN"],
"normStd": args["NORMALIZATION_STD"],
"vf": vf
}
preprocess_sample(sampleFile, params)
#converting the data sample into appropriate tensors for input to the model
visualFeaturesFile = sampleFile + ".npy"
videoParams = {"videoFPS": args["VIDEO_FPS"]}
inp, trgt, inpLen, trgtLen = prepare_main_input(
visualFeaturesFile, targetFile, args["MAIN_REQ_INPUT_LENGTH"],
args["CHAR_TO_INDEX"], videoParams)
inputBatch, targetBatch, inputLenBatch, targetLenBatch = collate_fn([
(inp, trgt, inpLen, trgtLen)
])
#running the model
inputBatch, targetBatch = (inputBatch.float()).to(device), (
targetBatch.int()).to(device)
inputLenBatch, targetLenBatch = (inputLenBatch.int()).to(device), (
targetLenBatch.int()).to(device)
model.eval()
with torch.no_grad():
outputBatch = model(inputBatch)
#obtaining the prediction using CTC deocder
if args["TEST_DEMO_DECODING"] == "greedy":
predictionBatch, predictionLenBatch = ctc_greedy_decode(
outputBatch, inputLenBatch, args["CHAR_TO_INDEX"]["<EOS>"])
elif args["TEST_DEMO_DECODING"] == "search":
beamSearchParams = {
"beamWidth": args["BEAM_WIDTH"],
"alpha": args["LM_WEIGHT_ALPHA"],
"beta": args["LENGTH_PENALTY_BETA"],
"threshProb": args["THRESH_PROBABILITY"]
}
predictionBatch, predictionLenBatch = ctc_search_decode(
outputBatch, inputLenBatch, beamSearchParams,
args["CHAR_TO_INDEX"][" "], args["CHAR_TO_INDEX"]["<EOS>"], lm)
else:
print("Invalid Decode Scheme")
exit()
#comiputing CER and WER
#cer = compute_cer(predictionBatch, targetBatch, predictionLenBatch, targetLenBatch)
#wer = compute_wer(predictionBatch, targetBatch, predictionLenBatch, targetLenBatch, args["CHAR_TO_INDEX"][" "])
#converting character indices back to characters
pred = predictionBatch[:][:-1]
pred = "".join([args["INDEX_TO_CHAR"][ix] for ix in pred.tolist()])
#printing the predictions
print("Prediction: %s" % (pred))
print("\n")